ENGR-E 533 Deep Learning Systems - Assignment 2

Name : Akhila Sakiramolla (asakiram@iu.edu)

UID : 2000886005

Importing required libraries

import numpy as np
import pandas as pd
import librosa
import IPython.display as ipd
import soundfile as sf
import tensorflow as tf
import keras
from tensorflow.keras import initializers
from keras.models import Sequential
from keras.models import Model
from keras.layers import Dense
from keras.layers.core import Dense, Activation
from tensorflow.keras.layers import Conv1D
from tensorflow.keras.layers import MaxPooling1D
from tensorflow.keras.layers import Flatten
from tensorflow.keras.layers import Conv2D
from tensorflow.keras.layers import MaxPooling2D
from tensorflow.keras.layers import Input

# To execute a cell line by line
from IPython.core.interactiveshell import InteractiveShell
InteractiveShell.ast_node_interactivity = "all"

Problem 1: Speech Denoising Using Deep Learning

# Loading data 

s, sr=librosa.load('train_clean_male.wav', sr=None)
sn, sr=librosa.load('train_dirty_male.wav', sr=None)

# Creating train and validation split for clean signal

s_train = s[0:1007119]
s_val = s[-251780:]

# Creating train and validation split for noisy signal

sn_train = sn[0:1007119]
sn_val = sn[-251780:]

# Coverting clean signals into spectograms

S_train = librosa.stft(s_train, n_fft=1024, hop_length=512)
S_val = librosa.stft(s_val, n_fft=1024, hop_length=512)

S_train_abs  = np.transpose(np.abs(S_train))
S_train_abs.shape

S_val_abs  = np.transpose(np.abs(S_val))
S_val_abs.shape

(1968, 513)

(492, 513)

# Coverting noisy signals into spectograms

X_train = librosa.stft(sn_train, n_fft=1024, hop_length=512)
X_val = librosa.stft(sn_val, n_fft=1024, hop_length=512)

X_train_abs  = np.transpose(np.abs(X_train))
X_train_abs.shape

X_val_abs  = np.transpose(np.abs(X_val))
X_val_abs.shape

(1968, 513)

(492, 513)

# Building a model and adding 3 hidden layers with 512 hidden units

model = Sequential()
model.add(Dense(512, input_dim = 513, activation='relu',kernel_initializer=tf.keras.initializers.HeNormal()))
model.add(Dense(512, activation='relu', kernel_initializer=tf.keras.initializers.HeNormal()))
model.add(Dense(512, activation='relu', kernel_initializer=tf.keras.initializers.HeNormal()))
model.add(Dense(513, activation='relu', kernel_initializer=tf.keras.initializers.HeNormal()))

# Calculating loss and optimizing it

model.compile(optimizer='adam',
              loss='mse',
              metrics=['accuracy'])

# Fitting the model on train dataset

model.fit(X_train_abs, S_train_abs, epochs=100, batch_size=128, shuffle=True)

Epoch 1/100
16/16 [==============================] - 1s 3ms/step - loss: 0.0850 - accuracy: 0.0523
Epoch 2/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0410 - accuracy: 0.1657
Epoch 3/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0253 - accuracy: 0.2642
Epoch 4/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0196 - accuracy: 0.3171
Epoch 5/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0164 - accuracy: 0.3664
Epoch 6/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0141 - accuracy: 0.3816
Epoch 7/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0127 - accuracy: 0.3913
Epoch 8/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0120 - accuracy: 0.3958
Epoch 9/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0109 - accuracy: 0.4085
Epoch 10/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0101 - accuracy: 0.4126
Epoch 11/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0095 - accuracy: 0.4284
Epoch 12/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0091 - accuracy: 0.4380
Epoch 13/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0086 - accuracy: 0.4380
Epoch 14/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0082 - accuracy: 0.4446
Epoch 15/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0083 - accuracy: 0.4466
Epoch 16/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0079 - accuracy: 0.4522
Epoch 17/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0080 - accuracy: 0.4502
Epoch 18/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0083 - accuracy: 0.4451
Epoch 19/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0081 - accuracy: 0.4477
Epoch 20/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0075 - accuracy: 0.4517
Epoch 21/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0071 - accuracy: 0.4527
Epoch 22/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0068 - accuracy: 0.4660
Epoch 23/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0066 - accuracy: 0.4736
Epoch 24/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0065 - accuracy: 0.4715
Epoch 25/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0062 - accuracy: 0.4675
Epoch 26/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0061 - accuracy: 0.4705
Epoch 27/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0058 - accuracy: 0.4756
Epoch 28/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0056 - accuracy: 0.4771
Epoch 29/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0054 - accuracy: 0.4807
Epoch 30/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0053 - accuracy: 0.4817
Epoch 31/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0052 - accuracy: 0.4868
Epoch 32/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0051 - accuracy: 0.4888
Epoch 33/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0050 - accuracy: 0.4792
Epoch 34/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0050 - accuracy: 0.4924
Epoch 35/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0049 - accuracy: 0.4827
Epoch 36/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0049 - accuracy: 0.4924
Epoch 37/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0047 - accuracy: 0.4858
Epoch 38/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0047 - accuracy: 0.4939
Epoch 39/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0047 - accuracy: 0.4980
Epoch 40/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0047 - accuracy: 0.4934
Epoch 41/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0047 - accuracy: 0.4909
Epoch 42/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0047 - accuracy: 0.4909
Epoch 43/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0048 - accuracy: 0.4939
Epoch 44/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0047 - accuracy: 0.4924
Epoch 45/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0046 - accuracy: 0.4985
Epoch 46/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0046 - accuracy: 0.4985
Epoch 47/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0046 - accuracy: 0.4990
Epoch 48/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0045 - accuracy: 0.4985
Epoch 49/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0043 - accuracy: 0.5020
Epoch 50/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0044 - accuracy: 0.5051
Epoch 51/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0044 - accuracy: 0.5041
Epoch 52/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0044 - accuracy: 0.5000
Epoch 53/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0048 - accuracy: 0.4970
Epoch 54/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0048 - accuracy: 0.5046
Epoch 55/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0049 - accuracy: 0.4903
Epoch 56/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0048 - accuracy: 0.4898
Epoch 57/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0045 - accuracy: 0.5041
Epoch 58/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0041 - accuracy: 0.5076
Epoch 59/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0041 - accuracy: 0.5071
Epoch 60/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0040 - accuracy: 0.5132
Epoch 61/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0040 - accuracy: 0.5066
Epoch 62/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0039 - accuracy: 0.5168
Epoch 63/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0038 - accuracy: 0.5193
Epoch 64/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0037 - accuracy: 0.5218
Epoch 65/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0036 - accuracy: 0.5224
Epoch 66/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0037 - accuracy: 0.5168
Epoch 67/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0036 - accuracy: 0.5305
Epoch 68/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0036 - accuracy: 0.5158
Epoch 69/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0036 - accuracy: 0.5315
Epoch 70/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0035 - accuracy: 0.5218
Epoch 71/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0035 - accuracy: 0.5285
Epoch 72/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0037 - accuracy: 0.5264
Epoch 73/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0040 - accuracy: 0.5249
Epoch 74/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0038 - accuracy: 0.5264
Epoch 75/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0039 - accuracy: 0.5264
Epoch 76/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0041 - accuracy: 0.5142
Epoch 77/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0044 - accuracy: 0.5076
Epoch 78/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0041 - accuracy: 0.5158
Epoch 79/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0039 - accuracy: 0.5183
Epoch 80/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0036 - accuracy: 0.5198
Epoch 81/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0034 - accuracy: 0.5203
Epoch 82/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0035 - accuracy: 0.5310
Epoch 83/100
16/16 [==============================] - 0s 5ms/step - loss: 0.0034 - accuracy: 0.5346
Epoch 84/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0032 - accuracy: 0.5381
Epoch 85/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0031 - accuracy: 0.5488
Epoch 86/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0031 - accuracy: 0.5396
Epoch 87/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0030 - accuracy: 0.5473
Epoch 88/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0029 - accuracy: 0.5498
Epoch 89/100
16/16 [==============================] - 0s 5ms/step - loss: 0.0029 - accuracy: 0.5574
Epoch 90/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0029 - accuracy: 0.5539
Epoch 91/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0031 - accuracy: 0.5513
Epoch 92/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0031 - accuracy: 0.5549
Epoch 93/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0031 - accuracy: 0.5473
Epoch 94/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0031 - accuracy: 0.5539
Epoch 95/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0030 - accuracy: 0.5534
Epoch 96/100
16/16 [==============================] - 0s 5ms/step - loss: 0.0030 - accuracy: 0.5539
Epoch 97/100
16/16 [==============================] - 0s 3ms/step - loss: 0.0029 - accuracy: 0.5518
Epoch 98/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0030 - accuracy: 0.5518
Epoch 99/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0029 - accuracy: 0.5579
Epoch 100/100
16/16 [==============================] - 0s 4ms/step - loss: 0.0029 - accuracy: 0.5549

<keras.callbacks.History at 0x7f9d10461590>

# Predicting on validation data 

S_test = model.predict(X_val_abs)

16/16 [==============================] - 0s 2ms/step

# Recovering the signal from predicted output

S_hat = np.multiply(X_val/X_val_abs.T,S_test.T)
S_rec = librosa.istft(S_hat, hop_length=512)
ipd.Audio(S_rec, rate = sr)

# Clipping clean signal to calculate SNR

s_val = abs(s_val[0:251392])
s_val.shape

(251392,)

# Calculating SNR value

SNR = 10*np.log10((np.sum(np.square(abs(s_val))))/(np.sum(np.square(abs(s_val) - abs(S_rec)))))
print("The SNR value for recovered signal : ", round(SNR,2))

The SNR value for recovered signal :  10.45

# Loading first test data and applying STFT

t1, sr = librosa.load('test_x_01.wav', sr=None)
T1 = librosa.stft(t1, n_fft=1024, hop_length=512)
T1_abs = np.abs(T1)
T1_abs = T1_abs.T
T1_abs.shape

(142, 513)

# Predicting on test data 
S_test1 = model.predict(T1_abs)

# Recovering the signal from predicted output
S_hat1 = np.multiply(T1/T1_abs.T,S_test1.T)
S_rec1 = librosa.istft(S_hat1, hop_length=512)
ipd.Audio(S_rec1, rate = sr)
sf.write('test_s_01_recons.wav', S_rec1, sr)

5/5 [==============================] - 0s 2ms/step

# Loading second test data and applying STFT

t2, sr = librosa.load('test_x_02.wav', sr=None)
T2 = librosa.stft(t2, n_fft=1024, hop_length=512)
T2_abs = np.abs(T2)
T2_abs = T2_abs.T
T2_abs.shape

(380, 513)

# Predicting on test data 
S_test2 = model.predict(T2_abs)

# Recovering the signal from predicted output
S_hat2 = np.multiply(T2/T2_abs.T,S_test2.T)
S_rec2 = librosa.istft(S_hat2, hop_length=512)
ipd.Audio(S_rec2, rate = sr)
sf.write('test_s_02_recons.wav', S_rec2, sr)

12/12 [==============================] - 0s 1ms/step

Problem 2: Speech Denoising Using 1D CNN

# Loading data 

s, sr=librosa.load('train_clean_male.wav', sr=None)
sn, sr=librosa.load('train_dirty_male.wav', sr=None)

# Creating train and validation split for clean signal

s_train = s[0:1007119]
s_val = s[-251780:]

# Creating train and validation split for noisy signal

sn_train = sn[0:1007119]
sn_val = sn[-251780:]

# Coverting clean signals into spectograms

S_train = librosa.stft(s_train, n_fft=1024, hop_length=512)
S_val = librosa.stft(s_val, n_fft=1024, hop_length=512)

S_train_abs  = np.transpose(np.abs(S_train))
S_train_abs.shape

S_val_abs  = np.transpose(np.abs(S_val))
S_val_abs.shape

(1968, 513)

(492, 513)

# Coverting noisy signals into spectograms

X_train = librosa.stft(sn_train, n_fft=1024, hop_length=512)
X_val = librosa.stft(sn_val, n_fft=1024, hop_length=512)

X_train_abs  = np.transpose(np.abs(X_train))
X_train_abs.shape

X_val_abs  = np.transpose(np.abs(X_val))
X_val_abs.shape

(1968, 513)

(492, 513)

# Building a model with 2 1D conloutional layers, 2 max pooling layers and 1 dense layer

model = Sequential()
model.add(Conv1D(filters=64, kernel_size=3, activation='relu',input_shape=(513,1)))
model.add(MaxPooling1D(pool_size=2))
model.add(Conv1D(filters=64, kernel_size=3, activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(Dense(513, activation='relu'))

# Calculating loss and optimizing it

model.compile(optimizer='adam',
              loss='mse',
              metrics=['accuracy'])

# Fitting the model

model.fit(X_train_abs, S_train_abs, epochs=100, batch_size=128, shuffle=True)

Epoch 1/100
16/16 [==============================] - 3s 41ms/step - loss: 0.0714 - accuracy: 0.0412
Epoch 2/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0430 - accuracy: 0.1550
Epoch 3/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0252 - accuracy: 0.2597
Epoch 4/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0169 - accuracy: 0.3283
Epoch 5/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0134 - accuracy: 0.3720
Epoch 6/100
16/16 [==============================] - 0s 7ms/step - loss: 0.0111 - accuracy: 0.3968
Epoch 7/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0097 - accuracy: 0.4060
Epoch 8/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0087 - accuracy: 0.4192
Epoch 9/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0084 - accuracy: 0.4309
Epoch 10/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0079 - accuracy: 0.4360
Epoch 11/100
16/16 [==============================] - 0s 7ms/step - loss: 0.0075 - accuracy: 0.4553
Epoch 12/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0069 - accuracy: 0.4522
Epoch 13/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0066 - accuracy: 0.4599
Epoch 14/100
16/16 [==============================] - 0s 7ms/step - loss: 0.0063 - accuracy: 0.4695
Epoch 15/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0059 - accuracy: 0.4812
Epoch 16/100
16/16 [==============================] - 0s 7ms/step - loss: 0.0056 - accuracy: 0.4812
Epoch 17/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0053 - accuracy: 0.4888
Epoch 18/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0051 - accuracy: 0.4909
Epoch 19/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0049 - accuracy: 0.5005
Epoch 20/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0048 - accuracy: 0.4970
Epoch 21/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0047 - accuracy: 0.5041
Epoch 22/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0045 - accuracy: 0.5107
Epoch 23/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0043 - accuracy: 0.5076
Epoch 24/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0042 - accuracy: 0.5122
Epoch 25/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0041 - accuracy: 0.5178
Epoch 26/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0041 - accuracy: 0.5163
Epoch 27/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0040 - accuracy: 0.5264
Epoch 28/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0038 - accuracy: 0.5239
Epoch 29/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0040 - accuracy: 0.5249
Epoch 30/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0038 - accuracy: 0.5305
Epoch 31/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0036 - accuracy: 0.5396
Epoch 32/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0035 - accuracy: 0.5346
Epoch 33/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0035 - accuracy: 0.5417
Epoch 34/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0033 - accuracy: 0.5432
Epoch 35/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0036 - accuracy: 0.5462
Epoch 36/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0035 - accuracy: 0.5366
Epoch 37/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0033 - accuracy: 0.5351
Epoch 38/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0033 - accuracy: 0.5407
Epoch 39/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0032 - accuracy: 0.5534
Epoch 40/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0032 - accuracy: 0.5513
Epoch 41/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0030 - accuracy: 0.5569
Epoch 42/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0029 - accuracy: 0.5523
Epoch 43/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0029 - accuracy: 0.5620
Epoch 44/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0029 - accuracy: 0.5691
Epoch 45/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0028 - accuracy: 0.5610
Epoch 46/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0027 - accuracy: 0.5706
Epoch 47/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0027 - accuracy: 0.5742
Epoch 48/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0026 - accuracy: 0.5732
Epoch 49/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0027 - accuracy: 0.5686
Epoch 50/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0026 - accuracy: 0.5666
Epoch 51/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0026 - accuracy: 0.5737
Epoch 52/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0029 - accuracy: 0.5727
Epoch 53/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0029 - accuracy: 0.5737
Epoch 54/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0027 - accuracy: 0.5661
Epoch 55/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0028 - accuracy: 0.5716
Epoch 56/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0027 - accuracy: 0.5828
Epoch 57/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0026 - accuracy: 0.5828
Epoch 58/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0025 - accuracy: 0.5854
Epoch 59/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0025 - accuracy: 0.5757
Epoch 60/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0024 - accuracy: 0.5833
Epoch 61/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0026 - accuracy: 0.5869
Epoch 62/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0026 - accuracy: 0.5879
Epoch 63/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0028 - accuracy: 0.5818
Epoch 64/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0025 - accuracy: 0.5828
Epoch 65/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0025 - accuracy: 0.5838
Epoch 66/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0024 - accuracy: 0.5950
Epoch 67/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0023 - accuracy: 0.5925
Epoch 68/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0025 - accuracy: 0.6006
Epoch 69/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0024 - accuracy: 0.5996
Epoch 70/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0025 - accuracy: 0.5879
Epoch 71/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0024 - accuracy: 0.5976
Epoch 72/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0023 - accuracy: 0.5971
Epoch 73/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0021 - accuracy: 0.6026
Epoch 74/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6042
Epoch 75/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6087
Epoch 76/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6087
Epoch 77/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6092
Epoch 78/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0022 - accuracy: 0.6062
Epoch 79/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0021 - accuracy: 0.6098
Epoch 80/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6026
Epoch 81/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6072
Epoch 82/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0022 - accuracy: 0.6016
Epoch 83/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0039 - accuracy: 0.5808
Epoch 84/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0039 - accuracy: 0.5655
Epoch 85/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0030 - accuracy: 0.5843
Epoch 86/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0027 - accuracy: 0.5884
Epoch 87/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0023 - accuracy: 0.5981
Epoch 88/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6037
Epoch 89/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0019 - accuracy: 0.6113
Epoch 90/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0018 - accuracy: 0.6143
Epoch 91/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0018 - accuracy: 0.6113
Epoch 92/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0019 - accuracy: 0.6225
Epoch 93/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0019 - accuracy: 0.6184
Epoch 94/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0018 - accuracy: 0.6199
Epoch 95/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0019 - accuracy: 0.6194
Epoch 96/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6138
Epoch 97/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0022 - accuracy: 0.6174
Epoch 98/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0020 - accuracy: 0.6113
Epoch 99/100
16/16 [==============================] - 0s 8ms/step - loss: 0.0017 - accuracy: 0.6098
Epoch 100/100
16/16 [==============================] - 0s 7ms/step - loss: 0.0015 - accuracy: 0.6270

<keras.callbacks.History at 0x7f9ca42e4e90>

# Predicting on validation data 

S_test = model.predict(X_val_abs)

16/16 [==============================] - 0s 9ms/step

# Recovering the signal from predicted output

S_hat = np.multiply(X_val/X_val_abs.T,S_test.T)
S_rec = librosa.istft(S_hat, hop_length=512)
ipd.Audio(S_rec, rate = sr)

S_rec.shape

(251392,)

# Clipping clean signal to calculate SNR

s_val = abs(s_val[0:251392])
s_val.shape

(251392,)

# Calculating SNR value

SNR = 10*np.log10((np.sum(np.square(abs(s_val))))/(np.sum(np.square(abs(s_val) - abs(S_rec)))))
print("The SNR value for recovered signal : ", round(SNR,2))

The SNR value for recovered signal :  12.34

# Loading first test data and applying STFT

t1, sr = librosa.load('test_x_01.wav', sr=None)
T1 = librosa.stft(t1, n_fft=1024, hop_length=512)
T1_abs = np.abs(T1)
T1_abs = T1_abs.T
T1_abs.shape

(142, 513)

# Predicting on test data 

S_test1 = model.predict(T1_abs)

5/5 [==============================] - 0s 31ms/step

# Recovering the signal from predicted output

S_hat1 = np.multiply(T1/T1_abs.T,S_test1.T)
S_rec1 = librosa.istft(S_hat1, hop_length=512)
ipd.Audio(S_rec1, rate = sr)
sf.write('test_s_01_recons_1Dcnn.wav', S_rec1, sr)

# Loading second test data and applying STFT

t2, sr = librosa.load('test_x_02.wav', sr=None)
T2 = librosa.stft(t2, n_fft=1024, hop_length=512)
T2_abs = np.abs(T2)
T2_abs = T2_abs.T
T2_abs.shape

(380, 513)

# Predicting on test data 

S_test2 = model.predict(T2_abs)

12/12 [==============================] - 0s 15ms/step

# Recovering the signal from predicted output

S_hat2 = np.multiply(T2/T2_abs.T,S_test2.T)
S_rec2 = librosa.istft(S_hat2, hop_length=512)
ipd.Audio(S_rec2, rate = sr)
sf.write('test_s_02_recons_1Dcnn.wav', S_rec2, sr)

Problem 3: Speech Denoising Using 2D CNN

# Loading data 

s, sr=librosa.load('train_clean_male.wav', sr=None)
sn, sr=librosa.load('train_dirty_male.wav', sr=None)

# Creating train and validation split for clean signal

s_train = s[0:1007119]
s_val = s[-251780:]

# Creating train and validation split for noisy signal

sn_train = sn[0:1007119]
sn_val = sn[-251780:]

# Coverting clean signals into spectograms

S_train = librosa.stft(s_train, n_fft=1024, hop_length=512)
S_val = librosa.stft(s_val, n_fft=1024, hop_length=512)

S_train_abs  = np.transpose(np.abs(S_train))
S_train_abs.shape

S_val_abs  = np.transpose(np.abs(S_val))
S_val_abs.shape

(1968, 513)

(492, 513)

# Coverting noisy signals into spectograms

X_train = librosa.stft(sn_train, n_fft=1024, hop_length=512)
X_val = librosa.stft(sn_val, n_fft=1024, hop_length=512)

X_train_abs  = np.transpose(np.abs(X_train))
X_train_abs.shape

X_val_abs  = np.transpose(np.abs(X_val))
X_val_abs.shape

(1968, 513)

(492, 513)

w = 20
train_data = X_train_abs[:w,:]

h = 1
while h <= X_train_abs.shape[0]-w:
  train_data = np.append(train_data,X_train_abs[h:h+w,:],axis=0)
  h += 1

x_train = train_data.reshape((1949,20,513,1))
x_train.shape

y_train = S_train_abs[:1949,:]
y_train = y_train.reshape((1949,513,1))
y_train.shape

(1949, 20, 513, 1)

(1949, 513, 1)

w = 20
val_data = X_val_abs[:w,:]

h = 1
while h <= X_val_abs.shape[0]-w:
  val_data = np.append(val_data,X_val_abs[h:h+w,:],axis=0)
  h += 1

x_val = val_data.reshape((473,20,513,1))
x_val.shape

(473, 20, 513, 1)

# Building a model with 2 22D conloutional layers and 2 max pooling layers 

model = Sequential()
model.add(Input(shape = (20,513,1)))
model.add(Conv2D(100, kernel_size=(7,7), activation='relu', strides = 1))
model.add(MaxPooling2D(pool_size=(2, 2), strides=2, padding="valid"))
model.add(Conv2D(50, kernel_size=(5,5), activation='relu',strides = 1))
model.add(MaxPooling2D(pool_size=(2, 2), strides=2, padding="valid"))
model.add(Flatten())
model.add(Dense(513, activation='relu',kernel_initializer = tf.keras.initializers.HeNormal()))

# Calculating loss and optimizing it

model.compile(optimizer='adam',loss='mse')

# Fitting the model

model.fit(x=x_train, y=y_train, batch_size=100, epochs=130, shuffle=True)

Epoch 1/130
20/20 [==============================] - 2s 81ms/step - loss: 0.0823
Epoch 2/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0665
Epoch 3/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0572
Epoch 4/130
20/20 [==============================] - 1s 52ms/step - loss: 0.0477
Epoch 5/130
20/20 [==============================] - 1s 52ms/step - loss: 0.0423
Epoch 6/130
20/20 [==============================] - 1s 52ms/step - loss: 0.0355
Epoch 7/130
20/20 [==============================] - 1s 52ms/step - loss: 0.0298
Epoch 8/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0255
Epoch 9/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0233
Epoch 10/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0211
Epoch 11/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0194
Epoch 12/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0174
Epoch 13/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0162
Epoch 14/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0154
Epoch 15/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0145
Epoch 16/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0137
Epoch 17/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0136
Epoch 18/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0131
Epoch 19/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0127
Epoch 20/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0120
Epoch 21/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0117
Epoch 22/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0114
Epoch 23/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0111
Epoch 24/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0108
Epoch 25/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0109
Epoch 26/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0106
Epoch 27/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0103
Epoch 28/130
20/20 [==============================] - 1s 55ms/step - loss: 0.0100
Epoch 29/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0098
Epoch 30/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0098
Epoch 31/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0099
Epoch 32/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0096
Epoch 33/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0097
Epoch 34/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0096
Epoch 35/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0095
Epoch 36/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0095
Epoch 37/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0091
Epoch 38/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0092
Epoch 39/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0089
Epoch 40/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0087
Epoch 41/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0086
Epoch 42/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0088
Epoch 43/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0093
Epoch 44/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0091
Epoch 45/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0087
Epoch 46/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0082
Epoch 47/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0081
Epoch 48/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0079
Epoch 49/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0079
Epoch 50/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0077
Epoch 51/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0077
Epoch 52/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0078
Epoch 53/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0080
Epoch 54/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0081
Epoch 55/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0079
Epoch 56/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0076
Epoch 57/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0075
Epoch 58/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0074
Epoch 59/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0076
Epoch 60/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0076
Epoch 61/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0076
Epoch 62/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0075
Epoch 63/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0077
Epoch 64/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0079
Epoch 65/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0078
Epoch 66/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0078
Epoch 67/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0076
Epoch 68/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0076
Epoch 69/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0075
Epoch 70/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0073
Epoch 71/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0073
Epoch 72/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 73/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 74/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 75/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 76/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 77/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0068
Epoch 78/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 79/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 80/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0070
Epoch 81/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 82/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 83/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0070
Epoch 84/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0072
Epoch 85/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0074
Epoch 86/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0072
Epoch 87/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0073
Epoch 88/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0071
Epoch 89/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 90/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0072
Epoch 91/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0074
Epoch 92/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0073
Epoch 93/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0070
Epoch 94/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 95/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0068
Epoch 96/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0067
Epoch 97/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 98/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 99/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0071
Epoch 100/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0071
Epoch 101/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0075
Epoch 102/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0073
Epoch 103/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 104/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0067
Epoch 105/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0065
Epoch 106/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0063
Epoch 107/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0065
Epoch 108/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0064
Epoch 109/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0065
Epoch 110/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0065
Epoch 111/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0064
Epoch 112/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0063
Epoch 113/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0065
Epoch 114/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0065
Epoch 115/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0064
Epoch 116/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0064
Epoch 117/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0066
Epoch 118/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0071
Epoch 119/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0072
Epoch 120/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0068
Epoch 121/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 122/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0068
Epoch 123/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0069
Epoch 124/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0068
Epoch 125/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0065
Epoch 126/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0064
Epoch 127/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0064
Epoch 128/130
20/20 [==============================] - 1s 53ms/step - loss: 0.0064
Epoch 129/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0063
Epoch 130/130
20/20 [==============================] - 1s 54ms/step - loss: 0.0062

<keras.callbacks.History at 0x7f9c9e293510>

# Predicting on validation data 

S_test = model.predict(x_val)

15/15 [==============================] - 0s 6ms/step

X_val1 = X_val[:,:473]
X_val_abs1 = X_val_abs[:473,:]

# Recovering the signal from predicted output

S_hat = np.multiply(X_val1/X_val_abs1.T,S_test.T)
S_rec = librosa.istft(S_hat, hop_length=512)
ipd.Audio(S_rec, rate = sr)

S_rec.shape

(241664,)

# Clipping clean signal to calculate SNR

s_val = abs(s_val[0:241664])
s_val.shape

(241664,)

# Calculating SNR value

SNR = 10*np.log10((np.sum(np.square(abs(s_val))))/(np.sum(np.square(abs(s_val) - abs(S_rec)))))
print("The SNR value for recovered signal : ", round(SNR,2))

The SNR value for recovered signal :  7.46

# Loading first test data and applying STFT

t1, sr = librosa.load('test_x_01.wav', sr=None)
T1 = librosa.stft(t1, n_fft=1024, hop_length=512)
T1_abs = np.abs(T1)
T1_abs = T1_abs.T
T1_abs.shape

(142, 513)

w = 20
train_data = T1_abs[:w,:]

h = 1
while h <= T1_abs.shape[0]-w:
  train_data = np.append(train_data,T1_abs[h:h+w,:],axis=0)
  h += 1

x_train = train_data.reshape((123,20,513,1))
x_train.shape

(123, 20, 513, 1)

# Predicting the model output

pred = model.predict(x_train,batch_size=32,verbose=0)

# Recovering the signal from predicted output

T1 = T1[:,:123]
T1_mag = (T1/np.abs(T1))
T1_rec = T1_mag*pred.T
T1_rec = librosa.istft(T1_rec,hop_length=512)
ipd.Audio(T1_rec, rate=sr)
sf.write('test_s_01_recons_2Dcnn.wav', T1_rec, sr)

# Loading second test data and applying STFT

t2, sr = librosa.load('test_x_02.wav', sr=None)
T2 = librosa.stft(t2, n_fft=1024, hop_length=512)
T2_abs = np.abs(T2)
T2_abs = T2_abs.T
T2_abs.shape

(380, 513)

w = 20
train_data = T2_abs[:w,:]

h = 1
while h <= T2_abs.shape[0]-w:
  train_data = np.append(train_data,T2_abs[h:h+w,:],axis=0)
  h += 1

x_train = train_data.reshape((361,20,513,1))
x_train.shape

(361, 20, 513, 1)

# Predicting the model output

pred = model.predict(x_train,batch_size=32,verbose=0)

# Recovering the signal from predicted output

T2 = T2[:,:361]
T2_mag = (T2/np.abs(T2))
T2_rec = T2_mag*pred.T
T2_rec = librosa.istft(T2_rec,hop_length=512)
ipd.Audio(T2_rec, rate=sr)
sf.write('test_s_02_recons_2Dcnn.wav', T2_rec, sr)